Prediction of OSNR with machine learning methods in DWDM/UDWDM long-haul transmission systems using EDFAs under the triple impact of FWM, SRS and ASE noise

In this research, four different machine learning (ML) algorithms have been used to predict the optical signal-to-noise ratio (OSNR) at central channels of dense and ultra-dense wavelength division multiplexing (DWDM/UDWDM) communication systems under the triple impact of four-wave mixing (FWM), stimulated Raman scattering (SRS), and amplified spontaneous emission (ASE) noise. The ML algorithms used in the research are Gaussian Process Regression (GPR), Narrow Neural Network (N-NN), Boosted Trees (BT), and Quadratic Support Vector Machines (QSVM). 7-, 15-, 31- and 63-channel configurations have been considered for DWDM/UDWDM systems. Two different datasets have been created to predict the OSNR value. The datasets used to predict the OSNR value have been created considering DWDM/UDWDM system configurations with 3.125 GHz-100 GHz channel spacing range and 1, 2, 4, and 5 erbium-doped fiber amplifiers (EDFAs). For the first data set, where the channel input powers have been varied between 0.1 mW and 5 mW, OSNR values have been predicted with the best goodness-of-fit metrics using GPR. For the second data set, where the total transmission length has been varied between 1 km and 150 km, OSNR values have been predicted with the best goodness-of-fit metrics using N-NN for 7-channel DWDM/UDWDM systems and GPR for 15-, 31-, and 63-channel DWDM/UDWDM systems. The detailed analysis presented in this paper will provide a new perspective on estimating the performance of AI integrated WDM-based transmission systems.